Vision-language reasoning on MathVista (test)
34.6AccuracyTop Probability + Confidence Modulation
Evaluation results
| Method | Links | |
|---|---|---|
| Top Probability + Confidence ModulationModel=LaViDa-Instruct, Decoding Method=Top Probability with Suffix Anchored Confidence Modulation (Full Method), Prompting Strategy=zero-shot2026.05 | 34.6 | |
| Top Margin + Confidence ModulationModel=LaViDa-Instruct, Decoding Method=Top Margin with Suffix Anchored Confidence Modulation (Full Method), Prompting Strategy=zero-shot2026.05 | 33.2 | |
| Top Margin + Suffix AnchorModel=LaViDa-Instruct, Decoding Method=Top Margin with Suffix Anchor, Prompting Strategy=zero-shot2026.05 | 29.2 | |
| Top Probability + Suffix AnchorModel=LaViDa-Instruct, Decoding Method=Top Probability with Suffix Anchor, Prompting Strategy=zero-shot2026.05 | 29.1 | |
| RandomModel=LaViDa-Instruct, Decoding Method=Random position selection, Prompting Strategy=zero-shot2026.05 | 28.8 | |
| Top ProbabilityModel=LaViDa-Instruct, Decoding Method=Top Probability (unmodified baseline), Prompting Strategy=zero-shot2026.05 | 27 | |
| Top MarginModel=LaViDa-Instruct, Decoding Method=Top Margin (unmodified baseline), Prompting Strategy=zero-shot2026.05 | 24.8 |